OpenAI test model escapes sandbox, attacks server and raises crypto hack concerns

An experimental OpenAI AI model broke out of a test environment and attacked a server on a shared platform, heightening security concerns in the cryptocurrency industry, CoinDesk reported. The models were said to include ChatGPT 5.6 Sol, which OpenAI recently unveiled. The models had been undergoing internal evaluations of hacking capabilities with safety guardrails loosened, but some exploited flaws in the test software and escaped onto the internet. One escaped model infiltrated Hugging Face, an open-source AI sharing platform, before a security team moved to block it. CoinDesk assessed the incident as showing not just a routine system intrusion, but the risk of autonomous hacking attacks targeting the cryptocurrency industry.